Abstract Ensuring fairness when training Machine Learning (ML) models remains a critical challenge, particularly when biases are embedded in the underlying data. This paper presents a fairness-aware graph structure learning framework demonstrating how learning fair graphs leads to fairer data for ML training and, consequently, fairer Artificial Intelligence (AI) decisioning based on such models. Our method incorporates a fairness regularization term into score-based structure learning algorithms, guiding the search towards graph structures that minimize discriminatory pathways while preserving statistical relationships. The learned fair graph structures enable the generation of synthetic datasets with mitigated biases, which can be used to train diverse ML models. This modification is non-trivial, as structure learning algorithms rely on local search strategies, while fairness is a global property that depends on the entire graph structure. Our framework is highly adaptable, compatible with various structure learning algorithms, and seamlessly incorporates different fairness metrics to meet specific contextual needs. Extensive experiments on both real-world and synthetic datasets demonstrate that our approach significantly improves fairness while maintaining competitive predictive performance, offering an interpretable and versatile solution for mitigating bias in AI systems..

Fraud and waste is a costly problem in medical insurance. Utilising sequence information for anomaly detection is under-explored in this domain. We present a multi-part method employing sequential pattern mining for identifying and grouping comparable courses of treatment, finding patterns within those courses, calculating the cost of possible additional or upcoded claims in unusual patterns, and ranking the providers based on potential recoverable costs. We applied this method to real-world radiation therapy data. Results were assessed by experts at the Australian Government Department of Health, and were found to be interpretable and informative. Previously unknown anomalous claim patterns were discovered, and confirmation of a previously suspected anomalous claim pattern was also obtained. Outlying providers each claimed up to ${\$}$486,617.60 in potentially recoverable costs. Our method was able to identify anomalous claims as well as the patterns in which they were anomalous, making the results easily interpretable. The method is currently being implemented for another problem involving sequential data at the Department of Health.

Analysing and learning from sequentially ordered symbolic data is increasingly important in applications such as finance and biology. Finding patterns that show interesting behaviour, such as regularly repeating occurrences within a time interval, can provide useful insight. In this paper we address the problem of efficiently identifying such behaviours. Existing approaches often require a target period to be specified, which will limit possible patterns to those approximating the specified periodicity, such as daily, monthly, quarterly and so on. In this paper we extend one such approach, derived from frequent pattern mining, to operate without the need for user-specified periodicity. Our new algorithms can identify the time interval and periodicity, or frequency of occurrence, of all periodically occurring patterns within a certain used-specified tolerance. Experimental results of our implementation show that the new approach can identify many more patterns in a real-world financial dataset, while on other sequential datasets it finds similar numbers of patterns without significant reduction in efficiency compared to existing approaches. We also verify the algorithm’s ability to recover recurring patterns in controlled experiments on synthetic data.

Our interest in this paper is in the construction of symbolic explanations for predictions made by a deep neural network. We will focus attention on deep relational machines (DRMs: a term introduced in Lodhi (2013)). A DRM is a deep network in which the input layer consists of Boolean-valued functions (features) that are defined in terms of relations provided as domain, or background, knowledge. Our DRMs differ from those in Lodhi (2013), which uses an Inductive Logic Programming (ILP) engine to first select features (we use random selections from a space of features that satisfies some approximate constraints on logical relevance and non-redundancy). But why do the DRMs predict what they do? One way of answering this was provided in recent work Ribeiro et al. (2016), by constructing readable proxies for a black-box predictor. The proxies are intended only to model the predictions of the black-box in local regions of the instance-space. But readability alone may not be enough: to be understandable, the local models must use relevant concepts in an meaningful manner. We investigate the use of a Bayes-like approach to identify logical proxies for local predictions of a DRM. As a preliminary step, we show that DRM's with our randomised propositionalization method achieve predictive performance that is comparable to the best reports in the ILP literature. Our principal results on logical explanations show: (a) Models in first-order logic can approximate the DRM's prediction closely in a small local region; and (b) Expert-provided relevance information can play the role of a prior to distinguish between logical explanations that perform equivalently on prediction alone.

This paper reports on the successful deployment of a peopleto-people recommender system in a large commercial online dating site. The deployment was the result of thorough evaluation and an online trial of a number of methods, including profile-based, collaborative filtering and hybrid algorithms. Results taken a few months after deployment show that key metrics generally hold their value or show an increase compared to the trial results, and that the recommender system delivered its projected benefits.

Users of online social networks such as dating websites often need help to find successful matches. People-to-people recommender systems can be used in social networks to help users find better matches, which requires solving the problem of reciprocal recommendation. However, most existing reciprocal recommenders use either profile similarity or interaction similarity to recommend new matches, without considering temporal features. In this paper we introduce a method for temporal reciprocal recommender systems using Hidden Markov Models to generate recommendations. Instead of summarising the whole historical data in one past state, we propose a model that formalises historical data on interactions as a series of successive states changing over time and then tries to find the recommended next state. We have implemented this new approach and the results of testing on industrial-scale data from a real dating website show a noticeable improvement over the previous best-performing recommenders.

This paper shows how to improve the recommendations of an interaction-based collaborative filtering (IBCF) recommender used in online dating. Previous work has shown that IBCF works well in this domain, although it tends to rank popular candidates highly, which leads to these users receiving a large number of contacts. We address this problem by using a Decision Tree model as a “critic” to re-rank the candidates generated by IBCF, effectively promoting less popular candidates. This method was first evaluated on historical data from a large online dating site and then trialled live on the same site by providing recommendations to a large number of users throughout a 9 week period. The live trial confirmed the consistency of the analysis on historical data and the ability of the method to generate suitable candidates over an extended period. Our recommendations gave higher success rates than those for a control group made with a baseline recommender.

Predicting people who other people may like has recently become an important task in many online social networks. Traditional collaborative filtering (CF) approaches are popular in recommender systems to effectively predict user preferences for items. One major problem in CF is computing similarity between users or items. Traditional CF methods often use heuristic methods to combine the ratings given to an item by similar users, which may not reflect the characteristics of the active user and can give unsatisfactory performance. In contrast to heuristic approaches we have developed CollabNet, a novel algorithm that uses gradient descent to learn the relative contributions of similar users or items to the ranking of recommendations produced by a recommender system, using weights to represent the contributions of similar users for each active user. We have applied CollabNet to the challenging problem of people to people recommendation in social networks, where people have a dual role as both "users" and "items", e.g., both initiating and receiving communications, to recommend other users to a given user, based on user similarity in terms of both taste (whom they like) and attractiveness (who likes them). Evaluation of CollabNet recommendations on datasets from a commercial online social network shows improved performance over standard CF.

MOTIVATION: Immunoglobulin heavy chain (IGH) genes in mature B lymphocytes are the result of recombination of IGHV, IGHD and IGHJ germline genes, followed by somatic mutation. The correct identification of the germline genes that make up a variable VH domain is essential to our understanding of the process of antibody diversity generation as well as to clinical investigations of some leukaemias and lymphomas. RESULTS: We have developed iHMMune-align, an alignment program that uses a hidden Markov model (HMM) to model the processes involved in human IGH gene rearrangement and maturation. The performance of iHMMune-align was compared to that of other immunoglobulin gene alignment utilities using both clonally related and randomly selected IGH sequences. This evaluation suggests that iHMMune-align provides a more accurate identification of component germline genes than other currently available IGH gene characterization programs. AVAILABILITY: iHMMune-align cross-platform Java executable and web interface are freely available to academic users and can be accessed at http://www.emi.unsw.edu.au/~ihmmune/.